{
  "model": "Qwen/Qwen3.5-0.8B",
  "revision": "2fc06364715b967f1860aea9cf38778875588b17",
  "lora_rank": 8,
  "effective_batch": 32,
  "weight_decay": 0.0,
  "max_grad_norm": 1.0,
  "stages": [
    {
      "name": "courtyard",
      "epochs": 5,
      "learning_rate": 5e-05,
      "examples": 779,
      "weighting": "equal trajectory",
      "loss_denominator": "actual_batch",
      "optimizer": "fresh"
    },
    {
      "name": "multilevel",
      "epochs": 4,
      "learning_rate": 1e-05,
      "examples": 4423,
      "weighting": "sqrt trajectory count by level, equal trajectory within level",
      "loss_denominator": "effective_batch",
      "optimizer": "fresh at stage start, retained between epochs"
    }
  ]
}